A Study On The Application Of Support Vector Machine In Advanced Control

This thesis studies the application of support vector machine (SVM) in the area of advanced process control. SVM, which has simple topological structure and good generalization capability, was put forward by Vapnik et. al. And it's a new and outstanding learning machine based on the statistical learning theory (SLT) and the structure risk minimization principle. The basic idea of SVM is to map linear inseparable input data into a high dimensional linear separable feature space via a nonlinear mapping technique (kernel dot product), and classification or regression is done in the feature space. SVM's solutions are characterized by convex quadratic optimization problems, which are typically solved in dual space. SVM has a lot of advantages in dealing with small samples, nonlinear system identification and control, etc.In detail, the major contributions of this thesis are as following:1. Research contributions and major problems in SLT and SVM are reviewed, mainly including the development and research state of SVM algorithms and their applications in system identification and control;2. In chapter 2, a new method for nonparametric model identification is proposed based on SVM with linear kernel function according to traditional pulse response and step response analyses. The presented method doesn't need special pulse response and step response tests. The nonparametric model can be gotten through the black-box identification method according to the ordinary daily data or random test data. And model algorithmic control and dynamic matrix control based on SVM with linearkernel function are put forward. Analytical solutions of control laws are obtained through predictive control mechanism;3. In chapter 3, one-step-ahead and multi-step-ahead predictive control structures and algorithms are proposed based on SVM with linear kernel function for weak nonlinear system. Analytical solutions of one-step-ahead and multi-step-ahead predictive control laws are obtained through predictive control mechanism;4. In chapter 4, predictive structure and algorithm based on SVM with quadratic polynomial kernel function for input-output nonlinear system are presented. According to test data or ordinary daily data, a predictive model is gotten through black-box identification. With feedback correcting and receding horizon optimization, the corresponding objective becomes a polynomial optimization problem with equality constraints of model outputs and bounded constraints of controller outputs. And the one-step-ahead analytical predictive controller output is easily gotten by using Cardan formula. For multi-step-ahead predictive problem, 2 cases are discussed. First, the special case of P=M (P is the prediction horizon and M is the control horizon) is studied, and the multi-step-ahead analytical predictive control law can be obtained by solving a series of cubic equations. Second, another case of P>M is also discussed, and the predictive control law is gotten through numeric optimization technique;5. New direct straight-model and inverse-model identification methods are developed by using SVM's excellent ability in function approximation. According to training data, linear and nonlinear systems' black-box identification is done using SVM with quadric polynomial and Gaussian RBF kernel functions respectively. And an internal model control method is put forward based on the SVM's inverse model and straight model;6. In chapter 6, a new kind of soft sensor is proposed based on SVM for microbiological fermentation process;7. Finally, a brief review of this thesis is given. Some future research directions are highlighted.